Ubisoft (Canada)

Offline Reinforcement Learning (RL) has emerged as a powerful alternative to imitation learning for behavior modeling in various domains, particularly in complex long range navigation tasks. An existing challenge with Offline RL is the signal-to-noise ratio, i.e. how to mitigate incorrect policy updates due to errors in value estimates. Towards this, multiple works have demonstrated the advantage of hierarchical offline RL methods, which decouples high-level path planning from low-level path following. In this work, we present a novel hierarchical transformer-based approach leveraging a learned quantizer of the state space to tackle long horizon navigation tasks. This quantization enables the training of a simpler zone-conditioned low-level policy and simplifies planning, which is reduced to discrete autoregressive prediction. Among other benefits, zone-level reasoning in planning enables explicit trajectory stitching rather than implicit stitching based on noisy value function estimates. By combining this transformer-based planner with recent advancements in offline RL, our proposed approach achieves state-of-the-art results in complex long-distance navigation environments.

Asynchronous waits are a common root cause of flaky tests and a major time-influential factor of Web application testing. We build a dataset of 49 reproducible asynchronous wait flaky tests and their fixes from 26 open source projects to study their characteristics in Web testing. Our study reveals that developers adjusted wait time to address asynchronous wait flakiness in about 63% of cases (31 out of 49), even when the underlying causes lie elsewhere. From this, we introduce TRaf , an automated time-based repair for asynchronous wait flakiness in Web applications. TRaf determines appropriate wait times for asynchronous calls in Web applications by analyzing code similarity and past change history. Its key insight is that efficient wait times can be inferred from the current or past codebase since developers tend to repeat similar mistakes. Our analysis shows that TRaf can statically suggest a shorter wait time to alleviate async wait flakiness immediately upon the detection, reducing test execution time by 11.1% compared to the timeout values initially chosen by developers. With optional dynamic tuning, TRaf can reduce the execution time by 16.8% in its initial refinement compared to developer-written patches and by 6.2% compared to the post-refinements of these original patches. Overall, we sent 16 pull requests from our dataset, each fixing one test, to the developers. So far, three have been accepted by the developers.

Learning to disentangle and represent factors of variation in data is an important problem in artificial intelligence. While many advances have been made to learn these representations, it is still unclear how to quantify disentanglement. While several metrics exist, little is known on their implicit assumptions, what they truly measure, and their limits. In consequence, it is difficult to interpret results when comparing different representations. In this work, we survey supervised disentanglement metrics and thoroughly analyze them. We propose a new taxonomy in which all metrics fall into one of the three families: intervention-based, predictor-based, and information-based. We conduct extensive experiments in which we isolate properties of disentangled representations, allowing stratified comparison along several axes. From our experiment results and analysis, we provide insights on relations between disentangled representation properties. Finally, we share guidelines on how to measure disentanglement.

Abstract The game audio designer Denis Zlobin uses his chapter to put forward the view that, rather than approach such design as the creation of sounds for video games, the designer should conceive of the role as creating video games via sound. To illustrate this, Zlobin uses the example of modulating player experience through the judicious use of intrinsically unpleasant sound effects. Zlobin proposes a framework for the evaluation, from various angles, of unpleasant sounds in video games then suggests how the audio designer can best create synergies with other areas of game development when seeking to use such sounds to serve the game.

To produce a video game, engineers and artists must iterate on the same project simultaneously. In such projects, a change to the work products of any of the teams can impact the work of other teams. As a result, any analytics tasks should consider intra- and inter-dependencies within and between artifacts produced by different teams. For instance, the focus of quality assurance teams on changes that are local to a team differs from one that impacts others. To extract and analyze such cross-disciplinary dependencies, we propose the multidisciplinary dependency graph. We instantiate our idea by developing tools that extract dependencies and construct the graph at Ubisoft---a multinational video game organization with more than 18,000 employees.

In online peer-to-peer games, players send periodic updates to each other and each player must locally reconstruct the position of their opponents in between these updates. In scenarios where players are driving cars, high speeds produce more pronounced errors in local replication of online opponents. In this work, we propose a new method of replicating opponents with less data sent and up to 45% less error compared to the state-of-the-art. We use a neural network-based approach to predict an opponent's position, combined with a path tracking controller from the field of mobile robotics, to produce smooth, believable trajectories for opponents' vehicles. We also propose a neural network-based approach to predict a replicated opponent's trajectory following a collision with a static obstacle.

Performance is key to the success and adoption of software systems. In video games, performance is commonly highlighted as one of the top quality concerns raised by players. To check the performance of their systems, development teams tend to rely on profiling and monitoring tools, which observe program executions to identify regressions. The usage of static analysis tools for this purpose has been so far limited. Lately, the success of Large Language Models in many code analytics tools led to attempts to leverage them in static performance analysis. These studies showed promising results in predicting runtime and regressions on large public datasets. In this paper, we evaluate the usability of such models in practice, and particularly in the domain of video games. We train a state-of-the-art neural network on the Code4Bench dataset to predict runtime regressions for programming competition programs, then evaluate its ability to generalize to new domains. Our results show that these models achieve great results (e.g. 95.73% accuracy for performance comparison) on the original domain for programs solving in-sample programming tasks, yet fail to generalize to out-of-sample tasks. Furthermore, we show that transfer techniques such as domain adversarial adaptation and model fine-tuning are not sufficient to transfer these models to the target industrial domain of AAA games.

We present our entry to the GENEA Challenge of 2022 on data-driven co-speech gesture generation. Our system is a neural network that generates gesture animation from an input audio file. The motion style generated by the model is extracted from an exemplar motion clip. Style is embedded in a latent space using a variational framework. This architecture allows for generating in styles unseen during training. Moreover, the probabilistic nature of our variational framework furthermore enables the generation of a variety of outputs given the same input, addressing the stochastic nature of gesture motion. The GENEA challenge evaluation showed that our model produces full-body motion with highly competitive levels of human-likeness.

The increasing complexity of gameplay mechanisms in modern video games is leading to the emergence of a wider range of ways to play games. The variety of possible play-styles needs to be anticipated and taken into account by designers, through automated tests. Reinforcement Learning (RL) is a promising answer to the need of automating video game testing. To that effect one needs to train an agent to play the game, while ensuring this agent will generate the same play-styles as the players in order to give meaningful feedback to the designers. We present CARMI : a Configurable Agent with Relative Metrics as Input. An agent able to emulate the players play-styles, even on previously unseen levels. Unlike current methods it does not rely on having full trajectories, but only summary data. Moreover it only requires little human data, thus compatible with the constraints of modern video game production. This novel agent could be used to investigate behaviors and balancing during the production of a video game with a realistic amount of training time.